Why a brighter measurement can be more precise
Repeated photographs of a stationary, uniformly illuminated scene do not produce exactly the same raw value. The number of detected photoelectrons fluctuates, and the electronics add their own error. We need a model before deciding whether to lengthen an exposure, average several frames, or smooth neighboring pixels. In particular, a smaller noise variance in an average does not by itself mean that the average preserves a sharp edge.
Definition 1.22.1
(Photosite observation model)
Work at one photosite and measure charge in electrons. Assume constant photon-generated electron rate , known exposure , and a dark-current electron rate . As a simplified independent-arrival model, let photoelectrons and dark electrons be independent. Let read noise have mean zero and variance in electrons squared, independently of both counts. A black-level-corrected, unsaturated raw observation is
The variance follows because a Poisson variable has variance equal to its mean and independent variances add. The stochastic dark count remains even if its mean is subtracted. This is an ideal model, not a promise that every camera has independent, constant-variance read noise. Fixed-pattern offsets, correlated electronics, digitization, saturation and changing light need separate treatment. Digital numbers must be converted to electrons with a calibrated gain before using these units.
Estimating the light rate
If is known, the rate can be estimated by subtracting the dark rate from the observed rate. The estimator is unbiased, and its variance has two terms with different dependence on .
Proposition 1.22.2
(Mean and variance of the rate estimator)
In the observation model above, let . Then
Proof
(Mean and variance by scaling)
By linearity of expectation, . Subtracting the constant does not change the variance, and . Hence
Proof X-ray
Where the two powers of exposure come from
The count variance grows in proportion to , while converting a count to a rate divides by . Variance therefore divides by , leaving the shot contribution proportional to . A read is incurred once with variance , so its rate variance has the faster factor . The derivation assumes no clipping and a constant rate throughout the exposure.
The first term comes from random arrival counts; the second from one read. Both fall as exposure increases under the stated conditions. The signal-to-noise ratio is defined next.
Definition 1.22.3
(Signal-to-noise ratio of the rate estimate)
When , the signal-to-noise ratio for estimating the photoelectron rate is
At this ratio is zero if noise is present; it is not a measure of accuracy for a noiseless zero signal, where the denominator vanishes.
Worked example 1.22.4
(One exposure)
Let , electrons per second, seconds and electrons. The expected raw charge is electrons and its variance is electrons squared. After subtracting the known dark rate, the rate estimate has mean and variance in squared electrons per second. Its standard deviation is electrons per second. The mean dark correction removes four electrons per second of bias, not the random dark-count fluctuations.
Common mistake
(Subtracting the mean dark level removes dark noise)
Subtracting from the observation removes the average dark contribution. It does not remove the Poisson fluctuation of the dark count , which still contributes to the variance. In the example above the variance stays after the subtraction; only the mean is corrected.
Worked example 1.22.5
(Which noise term dominates)
Take , and , so . The two terms are equal when , that is . At the variance is , of which comes from the read, and . At the variance is and . For short exposures the read term dominates, for long ones the shot term does. Edge case: if , then still falls with , but for every , because there is no signal to be resolved.
One long exposure or several short ones?
Suppose a fixed total observation time is divided into equal exposures, each of duration , with independent count and read noises, no gaps and the same rate. Sum their observations and subtract ; dividing by gives
The expected total count is unchanged by the split, so the total shot variance is unchanged. Each independent read adds another read-noise variance. Thus, within this model, one unsaturated exposure has no greater variance than a split exposure of the same total integration time. Short frames can still be useful to avoid clipping or motion blur, but then the simple comparison's assumptions change. A moving scene also makes direct pixelwise averaging biased for a desired reference frame unless images are aligned and occlusions handled.
Worked example 1.22.6
(The cost of extra reads)
For the preceding rates and , a single frame has rate variance . Four half-second frames have variance . Averaging their four separate rate estimates gives the same result. The four-frame standard deviation is , not half the single-frame standard deviation: the total integration time is fixed.
Counterexample mode
Averaging cannot undo clipping
Suppose a noiseless detector clips every charge above electrons to . Over two seconds a scene with rate produces electrons, and one with rate produces , yet both recorded values are . Repeating and averaging these clipped values still gives and cannot distinguish the two rates. The unbiased formula above explicitly excludes clipping; it must not be applied to saturated observations as if they were linear.
Spatial averaging and shared samples
For a fixed weighted linear estimate , its variance is
This identity does not require Gaussian noise. Four independent samples of the same underlying signal and equal variance have mean variance . If their signals differ, the same average estimates their mean; using it as an estimate of one center value can introduce bias. At edges, that bias is the price of blending unlike locations. With signal-dependent shot noise, equal variances are not automatic either.
Demosaicing creates dependencies even when the raw samples are independent. Let and , with independent equal-variance raw errors of variance . Then , but because the two estimates reuse . Treating reconstructed pixels as independent would miss this covariance. Subsequent color matrices and spatial filters propagate the full covariance, not just individual variances.
Common mistake
(Treating demosaiced neighbors as independent)
With and as above, independence would give . The shared sample gives , because has variance . Ignoring the covariance reports where the true value is , so the variance of any filter that combines both pixels is understated.
Practice with solutions
Quick check
(With , , and , find the mean and variance of and the variance of .)
Use the variance formula for and then divide by .
Solution · Answer
and ; dividing by leaves rate variance .
Quick check
(For the same rates and , compare the rate variances of one frame and two equal frames. State the assumptions.)
Apply the split-exposure formula with and .
Solution · Answer
One read gives ; two reads give . This assumes independent reads and arrivals, identical constant rate, no gaps and no saturation.
Quick check
(Two measurements of constant signal have errors of variance and covariance . Find the variance of their average.)
Use the weighted-sum identity with both weights equal to .
Solution · Answer
The variance is . Independence would instead give .
Quick check
(A two-pixel edge has true values and . A noiseless average of the two is used to estimate the first value. Find its error and explain why averaging can still be misleading without noise.)
Compute the average first and compare it with the first true value.
Solution · Answer
The average is , giving error relative to the first pixel. Averaging different signals can create deterministic bias even when random variance is zero.